freestar.queue = freestar.queue || []; freestar.config = freestar.config || {}; freestar.config.enabled_slots = []; freestar.queue.push(function() { googletag.pubads().set('page_url', ''); }); freestar.initCallback = function () { (freestar.config.enabled_slots.length === 0) ? freestar.initCallbackCalled = false : freestar.newAdSlots(freestar.config.enabled_slots) } The Game Haus - Sports & Esports news, from our Haus to yours.
The Game Haus
Dota 2 Esports

DOTA 2: The Evolution of OpenAI

The Evolution of OpenAI

It’s easy to quip about AI being the end of humanity, making references to Terminator all the while.

You’ll see that we’re not entirely above such basic humor later on, but for now, let’s introduce the evolution of OpenAI and bring you up to speed on its impact on the Dota 2 pro scene.

OpenAI is an artificial intelligence company founded by Sam Altman and SpaceX and Tesla CEO Elon Musk. Their general mission is to advance AI for the betterment of humanity. A noble aspiration to be sure, but what does this have to do with Dota 2?

We’ve got all the answers right here, so come with us if you want to live (told you…)

The Evolution of OpenAI

OpenAI began life in December of 2015. Their mission is to expand the capabilities of current AI systems and open the door for AI to become an extension of human will. When looking for a testing ground for this new AI system, the developers chose Dota 2.

The OpenAI team chose Dota 2 because of the high degree of complexity, the need for adaptation and the multitude of potential combinations of moves, actions and reactions. In order to prepare itself to take on the pros, OpenAI had to get smart.

To this end, OpenAI uses a reinforcement learning algorithm called Proximal Policy Optimization (PPO). If that sentence made your eyelids feel heavy don’t worry, here’s breakdown.

The Evolution of OpenAI
Image courtesy of OpenAI

PPO is essentially a method of reinforcing AI behavior through trial and error. The AI is presented with a task or a problem and through trial and error it works out the most efficient way to solve it. This allows the developers to introduce hazards and hindrances that the AI must overcome, learning all the while.

Never is this more evident than in this hilarious clip of an AI avatar attempting to reach an orb while being pelted by digital snowballs. The AI will utilize PPO to overcome the challenge and learn skills useful to its current environment. Skills like how to maintain its balance and how to stand up when it gets knocked down.

Introduction of OpenAI to Dota 2

To start with, OpenAI knew nothing about the game of Dota 2. When starting out the AI doesn’t know about last hitting or even the objective of the game. The AI doesn’t even know it’s playing Dota, it’s just attempting to solve a problem in the most efficient way possible.

OpenAI doesn’t have any concept of a UI, or what a hero or ability looks like. Its thinking is purely mathematical. All OpenAI ‘sees’ is a selection of numbers, with its objective being to simply optimize the numbers in its favor.

What OpenAI knows is that when it moves or casts an ability some numbers changed – numbers related to health, mana, hero position, creep behavior, gold, etc. It doesn’t know initially if the change is good or bad, it just knows something has changed. It then (through a very long process of trial and error) works out if the change is beneficial or detrimental.

The Evolution of OpenAI
Image courtesy of Valve

To learn a new patch, the developers take the latest version of OpenAI and drop it into the new patch. It notices that certain numbers change differently to how they did before, or perhaps don’t change at all. OpenAI will modify its behavior accordingly in favor of solving the puzzle (in this case winning a game of Dota 2).

OpenAI’s CTO Greg Brockman explained that the OpenAI bot was able to learn to play the game at a professional level from scratch in the span of just two weeks of real time (336 hours). Even more amazingly, after just 1 hour of training the OpenAI bot is able to crush the in-game bots. The evolution of OpenAI as a competitive gaming opponent is incredible.

OpenAI’s First Live Match

By August of 2017 OpenAI was ready to move into the big league. On August 11th 2017, during The International 17 (TI17) in Seattle in front of 20,000 people, OpenAI played a 1v1 mid-only game against the ever-charismatic Danil “Dendi” Ishutin.

The Evolution of OpenAI
Image courtesy of Natus Vincere

The match consisted of a best of 3 series, with both combatants using Shadow Fiend. In a hotly contested first match, OpenAI utilized incredible Shadowraze faking and masterful positioning to secure two kills without reply, winning the first game.

In the second, decisive game, Dendi realized almost immediately that his chances of winning were almost zero. He calls GG with less than 90 seconds played.

The Future

The evolution of OpenAI is fascinating to watch. It has gone from a general learning system to mastering a game as complex as Dota 2 in the space of two weeks. It is the tip of the iceberg for OpenAI as a system to enhance and build value for humanity.

AI is not without its risks, as OpenAI themselves have discussed, but with diligent professionals committed to improving AI such that it rivals human performance on almost every intellectual task, the future is certainly exciting for AI.

Tune in later in the week as we explore OpenAI’s recent performances in Dota 2 on a true 5v5 scale. While you’re here, why not check out some of our other great sports and esports articles at The Game Haus?

Follow us

You can like The Game Haus on Facebook and follow us on Twitter for more sports and esports articles

Follow Matt on Twitter @MattyMead2006.

From Our Haus to Yours.

Related posts

Frankfurt DOTA 2 Major Groups

The Game Haus Staff

Five Reasons To Love Your Bad Team

Robert Hanes

1v1 Me Bruh!

The Game Haus Staff

Thanks for reading! Let us know what your thoughts are on the article!

Share This